AI Photo Control
A computer-vision system replacing manual review of courier photos — built on a unit-economics case, designed human-in-the-loop from day one.
01 · context
Every courier must look the part and be who they claim: uniform, equipment, documents. At Uzum Tezkor's scale — 2,000+ active couriers a day — that means a continuous stream of photos that operators were reviewing by hand.
I drive this as a product: from “can AI do this cheaper and faster?” to a rollout the business and legal can live with.
02 · problem
Manual photo review doesn't scale. It is slow, repetitive, error-prone at the end of a shift, and its cost grows linearly with the courier base.
Worse, the interesting failures are rare: an operator sees far more clean photos than violations, which is exactly the setup where humans start rubber-stamping.
03 · process
Unit economics before models
Built the cost/ROI case first: modern vision-language models put AI screening at roughly cents per 1,000 photos, against an estimated $4–11 for the same thousand reviewed manually. That gap — not the tech novelty — is what sold the project. The numbers are a cost model from discovery, and I present them as such.
Human-in-the-loop by design
Designed the rollout so AI filters the obvious passes and flags the doubtful cases to a human. Trade-off accepted: lower automation percentage at the start in exchange for trust and a measurable error baseline before widening the gate.
Risks surfaced before they became blockers
Courier photos are personal data. I raised the privacy and legal questions early and built mitigations into the roadmap, instead of letting them surface as a launch-week veto.
Prove it, then scale it
The 2026 goal is explicit: get Photo Control to “proven in production” — measured accuracy, measured savings — not just a working demo.
04 · solution
A computer-vision verification pipeline for courier uniform and identity checks, with operators kept in the loop for flagged cases.
- Automated screening of courier photos (uniform / equipment / documents)
- Confidence-based routing: clear passes auto-approved, doubts go to a human
- Cost model comparing AI vs manual review per 1,000 photos
- Privacy and legal mitigations planned into the roadmap, not patched in
05 · results
In practice
- The project was green-lit on economics, not hype — the case survived finance scrutiny.
- Human-in-the-loop design earned buy-in from ops instead of resistance.
- Legal/privacy review happened on my initiative before anyone asked.
06 · learnings
- For AI projects, the spreadsheet convinces before the demo does.
- “Cost model” and “measured result” are different claims — mixing them up costs credibility exactly when you need it.
- Raise legal risks early enough and the lawyers end up co-authoring the rollout.
07 · routine ops
- SQL analysis of verification volumes and operator load
- Stakeholder alignment: ops, engineering, legal, finance
- Vendor/model cost tracking as prices move